Questions about SwinUNETR pre-training details on ~50,000 Volumes

[moyiliyi]

Hi, thank you for sharing this great work!

I noticed that in the tutorial "Pre-trained SwinUNETR Backbone on ~50,000 3D Volumes", it mentioned that the pre-training was done using ~50,000 3D volumes from 14 datasets.

However, when I checked the referenced paper "Disruptive Autoencoders: Leveraging Low-level features for 3D Medical Image Pre-training", it seems that only ~10,000 volumes were used for pre-training.

I have a few questions about this:

1. Are there any papers or technical reports that provide more details about the results using the ~50,000 3D volumes?
2. Did the inclusion of additional data lead to any performance improvements?
3. Which specific weight file corresponds to the pre-training with ~50,000 3D volumes?

I’d greatly appreciate any clarification or pointers to relevant resources.
Thank you for your time and clarification!